Active learning enabled reactor characterization for mass transfer in aerobic oxidation reactions

Abstract

Determination of mass transfer coefficients (k_La) plays a critical role in multiple biopharmaceutical operations ranging from aerobic oxidation reactions for small molecule process development to cell-culture based processes for large molecule process development. Accurate determination of k_La across different scales and reactor configurations is required to develop such processes in a robust and scalable manner. We report the development of a machine learning (ML) based model that accurately predicts k_La across different scales ranging for 100 mL to 100 L. We demonstrate that the ML model can further be used for predictive modeling, such as interpreting sensitivity and estimating impact of new process conditions. Furthermore, integrating the ability to estimate uncertainty in the model prediction, we developed a general framework for a diversified uncertainty-based active learning (AL) algorithm which iteratively recommends experiments based on information criteria and design-space exploration. The novel application of this framework enables automated mass transfer characterization of a previously unexplored reactor configuration. Specifically, we show that using this AL-guided iterative design of experiment led to progressive improvement in the model's forecasting accuracy, improving from 39% at the beginning of AL iterations to 90% at the completion of five AL iterations. Our results confirm that this AL framework offers an efficient closed-loop experimentation strategy that minimizes experimental burden required to accurately characterize mass transfer coefficients for aerobic oxidation processes.